Uncertainties

CANYON-B provides comprehensive uncertainty estimates for all predictions. These uncertainties are composed of multiple components and are provided as standard uncertainties (1σ).

Uncertainty Components

Measurement Uncertainty (_cim)

• Based on input measurement errors
• Configurable via input parameters
• Default values:

  epres = 0.5    # Pressure error (dbar)
  etemp = 0.005  # Temperature error (°C)
  epsal = 0.005  # Salinity error
  edoxy = None   # Oxygen error (defaults to 1% of value)
  

Neural Network Uncertainty (_cin)

• Derived from committee disagreement
• Represents model uncertainty
• Includes:
  • Committee variance
  • Bias terms
  • Network-specific uncertainties

Input Propagation Uncertainty (_cii)

• How input errors affect prediction
• Calculated using local sensitivity
• Based on error propagation theory

Total Uncertainty (_ci)

Combines all components, and provided as a standard uncertainty:

total_uncertainty = sqrt(cim² + cin² + cii²)

Accessing Uncertainties

results = canyonb(**data)

# Total uncertainty
ph_uncertainty = results['pH_ci']

# Component uncertainties
measurement_unc = results['pH_cim']
network_unc = results['pH_cin']
input_unc = results['pH_cii']

Parameter-Specific Considerations

Carbonate System

• pH: Additional term for scale conversion
• pCO2: Non-linear error propagation
• AT/CT: Fixed measurement uncertainty terms

Nutrients

• Relative errors increase at low concentrations
• Additional terms for seasonal variability
• Regional uncertainty considerations